Skip to content
noveltodrama

Why chapter splitting breaks: five manuscript formats that defeat naive parsers

第X章 headings inside dialogue, novels with no headings at all, 40,000-word chapters, bilingual files and EPUB dumps — what breaks automatic splitting and how to fix it before you pay for summaries.

noveltodrama editorial 6 min read

Every adaptation problem downstream of import is caused by a bad split. If chapter 7 is actually the last 400 words of chapter 6 plus the first 2,000 of chapter 7, the summary is wrong, the episode map is wrong, and the dialogue that gets pulled forward is from the wrong scene.

These are the five formats that break naive splitting, in the order we see them.

1. Headings that look like body text

Chinese web novels use 第X章, but so do characters. A line reading 她翻到第三章,愣住了 is prose, not a heading. A parser that matches 第.*章 anywhere in a line will split mid-scene.

The fix is positional, not textual: a heading is a line that stands alone, is short, and — in most published files — has a blank line before it. Our splitter requires the preceding line to be empty and caps heading length at 60 characters. It still occasionally over-splits on a scene break, which is why the preview exists.

2. No headings at all

Literary fiction and translated novels often use blank-line scene breaks instead of chapters. Some files are one continuous block of 80,000 words.

When fewer than two headings are found, the splitter falls back to word-count chunking: accumulate paragraphs until about 3,000 words, then cut. That is not a chapter — it is a slice — but it gives you units small enough to summarise and re-cut by hand.

If your file has real chapter breaks expressed only as *** or ◆, convert them to Chapter N before import. Two minutes of find-and-replace saves an hour of manual re-cutting.

3. Chapters longer than an episode can hold

A 9,000-word chapter is common at climaxes. It compresses to maybe 1,500 words of dialogue, which is 8–10 episodes. The splitter will keep it as one chapter because it is one chapter, and the episode map then has to subdivide it.

Subdividing is fine, but it should be explicit: give the chapter a note in the map (ch. 88 → episodes 19–22, confrontation) so the outline step knows to expand rather than compress. Summaries are per chapter, so you pay 3 credits once and get one summary that has to carry four episodes — a summary that thin will produce thin outlines.

4. Bilingual and mixed-script files

A file with Chinese prose and English chapter headings, or an English translation with Chinese titles retained, splits fine but summarises badly if the source locale is set wrong. Set the source locale explicitly at import; the summariser and the dialogue writer both use it, and a mismatch produces dialogue in the wrong language that then has to be re-run.

5. EPUB and DOCX dumps

Pasting from an EPUB reader usually brings soft line breaks at the column width, which destroys paragraph detection and makes every sentence look like a heading candidate. DOCX pastes carry list numbering that can match the 1. Title pattern.

Convert to plain text with hard-wrapped paragraphs first. If a line ends mid-sentence and the next line continues it, the file is hard-wrapped and needs re-flowing before import.

The check that costs nothing

Preview the split before importing. It is free, it does not store anything, and it shows you chapter count, per-chapter word counts and the estimated summary cost. If chapter 3 is 12 words and chapter 4 is 11,000, the split is wrong and no amount of prompting downstream will fix it.

Fix the file, not the prompt.

Your first episode starts with 300 free credits

Enough to import a short manuscript, summarise ten chapters and board one full episode. No card required.

300 credits · No card required · Commercial rights on paid plans